ICML 2026 Papers — Page 38
International Conference on Machine Learning · 6554 papers
Neural QAOA$^2$: Differentiable Joint Graph Partitioning and Parameter Initialization for Quantum Combinatorial Optimization
Zubin Zheng (Southern University of Science and Technology), Shengcai Liu (Southern University of Science and Technology)
OptimizationGraph Neural NetworkSupervised Fine-TuningGraphTabularBenchmarkPhysics Related
🎯 What it does: Propose Neural QAOA 2, a differentiable divide-and-conquer framework that jointly generates graph partitioning and QAOA parameters, addressing the partitioning metric mismatch and topology-unaware initialization issues in QAOA2.
Neural Quantum States in Mixed Precision
Massimo Solinas (University of Regensburg), Roeland Wiersema (Flatiron Institute)
Computational EfficiencyConvolutional Neural NetworkMixture of ExpertsTabularTime SeriesSequentialPhysics Related
🎯 What it does: This study explores the role of mixed-precision arithmetic in variational Monte Carlo (VMC) based on neural networks, particularly how low-precision formats can accelerate simulations of quantum many-body systems without compromising accuracy.
Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights
Yulu Gan (Massachusetts Institute of Technology), Phillip Isola (Massachusetts Institute of Technology)
OptimizationKnowledge DistillationRepresentation LearningHyperparameter SearchData-Centric LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningGaussian SplattingTextMultimodality
🎯 What it does: Investigated the structure of the parameter space around pre-trained models, discovering a large number of task experts near large models, and proposed the RandOpt algorithm, which combines random guessing and ensemble learning, using random perturbations to quickly locate and aggregate multi-task experts.
Neural Vector Lyapunov–Razumikhin Certificates for Delayed Interconnected Systems
Jingyuan Zhou (National University of Singapore), Kaidi Yang (National University of Singapore)
Autonomous DrivingOptimizationReinforcement Learning from Human FeedbackTabularTime SeriesBenchmarkPhysics Related
🎯 What it does: Proposes an scalable, delay-aware neural vector Lyapunov-Razumikhin proof framework for synthesizing and verifying stability proofs of large-scale delayed-coupled systems;
Neural-HSS: Hierarchical Semi-Separable Neural PDE Solver
Pietro Sittoni (Gran Sasso Science Institute), Francesco Tudisco (Gran Sasso Science Institute)
OptimizationComputational EfficiencyPoint CloudMeshTabularPhysics Related
🎯 What it does: This paper proposes a neural network architecture called Neural-HSS for efficiently learning solvers for partial differential equations (PDEs), specifically targeting the Green's function structure of elliptic PDEs;
Neural-Inspired Modeling of Auditory Selection and Compensation for Audio-Visual Speech Separation
Xinmeng Xu (Lingnan University), S. Joe Qin (Lingnan University)
RecognitionRestorationConvolutional Neural NetworkRecurrent Neural NetworkTransformerAuto EncoderContrastive LearningVideoMultimodalityAudio
🎯 What it does: Constructed Neuro-SCNet, a two-stage audio-visual speech separation framework based on neuroscience: first performing explicit auditory selection (audio gating), and then restoring suppressed speech information through cross-modal compensation (visual residual).
Neural–Evolutionary Symbolic Regression with Global Constraints: Constraint-Aware Decoding and Reward Shaping
Xiangdong Wu (Beihang University), Rongye Shi (Beihang University)
OptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningReinforcement Learning from Human FeedbackGraph Neural NetworkReinforcement LearningPrompt EngineeringGenerative Adversarial NetworkContrastive LearningGraphTabularBenchmarkPhysics Related
🎯 What it does: Propose a symbolic regression framework based on graph neural networks called GCN-SR, which explicitly maintains the expression tree structure using fixed-depth symbolic perfect binary trees (SPBT), and combines genetic programming (GP) for local refinement, employing a similarity-weighted strategy (SWPG) for reward shaping.
NeuralFLoC: Neural Flow-Based Joint Registration and Clustering of Functional Data
Xinyang Xiong (ShanghaiTech University), Pengcheng Zeng (ShanghaiTech University)
OptimizationRepresentation LearningConvolutional Neural NetworkFlow-based ModelContrastive LearningTime SeriesSequentialStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes an end-to-end unsupervised deep learning framework called NeuralFLoC for simultaneously performing diffeomorphic temporal alignment and clustering of functional data.
NeurIPS: Neuro-anatomical Inductive Priors for Sphere-based Brain Decoding
Sijin Yu (South China University Of Technology), Xin Zhang (South China University Of Technology)
RestorationTransformerMixture of ExpertsDiffusion modelAuto EncoderImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose the NeurIPS framework, achieving cross-subject image reconstruction through anatomy-prior-driven fMRI decoding on surface meshes.
Neuro-evolutionary Continual Reinforcement Learning
Pengyi Li (Tianjin University), Jianye HAO
OptimizationComputational EfficiencyReinforcement LearningScore-based ModelAuto Encoder
🎯 What it does: Propose the Nevo-CRL framework, which utilizes a fixed-capacity single policy network combined with sparse masks for continual reinforcement learning. It addresses issues of catastrophic forgetting, negative transfer, and parameter utilization by constructing masks based on semantic similarity, population evolution with importance-guided crossover, shared experience replay, and network pruning.
Neuro-Fuzzy Concept Learning for Interpretable Large Multimodal Models
Ritik Mishra (Indian Institute of Technology Indore), M. Tanveer (Indian Institute of Technology Indore)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelVision Language ModelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes the Neuro-FeX framework, which combines neural fuzzy inference (TSK) with semi-non-negative matrix factorization to extract interpretable multimodal concepts from large-scale multimodal models and generate IF-THEN rules.
Neuro-Symbolic AI for Analytical Solutions of Differential Equations
Orestis Oikonomou (ETH Zurich), Georgios Kissas (Swiss Data Science Center)
OptimizationTransformerAuto EncoderTabularTime SeriesBenchmarkPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes SIGS, a neural symbolic framework based on context-free grammar, for automatically discovering analytical solutions of PDEs.
NeuroCLUS: A Foundation Model with Functional Clustering for Intracranial Neural Decoding
Hui Zheng (Independent Researcher), Haiteng Wang
ClassificationRecognitionAnomaly DetectionConvolutional Neural NetworkTransformerSupervised Fine-TuningAuto EncoderContrastive LearningTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: Propose NeuroCLUS, a foundational model that learns through functional clustering, for unsupervised pre-training and multi-task decoding of electroencephalogram (EEG) signals.
NeurOCNN: A Neural-Operator-Based Model for Physiological Time Series
Daya Kumar (University of Western Ontario), Apurva Narayan (University of Western Ontario)
ClassificationConvolutional Neural NetworkTransformerContrastive LearningTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: Proposes NeurOCNN, a physiological time series model based on neural operators, continuous-time spline convolution, Fourier projection pooling, and attention heads, for achieving function-to-label mapping of multi-channel physiological signals.
NeuroMamba: A Universal Spatiotemporal Module for Robust Perception in Degraded Sensory Streams
Jinfeng Li (Hangzhou Dianzi University), Pan Li (Hangzhou Dianzi University)
RestorationObject DetectionAutonomous DrivingSpiking Neural NetworkTransformerSupervised Fine-TuningDiffusion modelContrastive LearningVideoPoint Cloud
🎯 What it does: Propose a pluggable module called NeuroMamba, used to restore spatial-temporal consistency and enhance perception robustness in severely degraded sensor streams.
Neuromem: A Granular Decomposition of the Streaming Lifecycle in External Memory for LLMs
Ruicheng Zhang (Huazhong University of Science and Technology), Hai Jin (Huazhong University of Science and Technology)
RetrievalCompressionOptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the Neuromem testing platform to conduct fine-grained evaluation of external memory modules in real-time streaming insertion and retrieval environments, decomposing the lifecycle into five design dimensions (data structure, normalization, merging, querying, and context fusion), and quantifying the impact of each dimension through cross-experiments;
NeuronCtrl: Geometry-Aware Safe Closed-Loop Generative Control for Neuronal Microenvironment Dynamics
Haowei Xu (Peking University), Zhaoheng Xie (Peking University)
OptimizationSafty and PrivacyRobotic IntelligenceRecurrent Neural NetworkGraph Neural NetworkReinforcement LearningDiffusion modelFlow-based ModelGenerative Adversarial NetworkGraphTime SeriesBiomedical DataBenchmark
🎯 What it does: This paper proposes a geometry-aware, safety-closed-loop generation control framework called NEURONCTRL, for real-time regulation of high-dimensional fields in neural microenvironments under sparse measurements.
NeurVLA: Unleashing Failure-Handling Capability of Vision-Language-Action Models via Neural-Symbolic Reasoning
Xuqi Liu (Zhejiang University), Siliang Tang (Zhejiang University)
Autonomous DrivingRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningVision-Language-Action ModelContrastive LearningImageTextMultimodalityChain-of-Thought
🎯 What it does: Propose NeurVLA, a neuro-symbolic framework that utilizes observation-driven failure correction and trajectory modeling for failure prevention, and internalizes these failure handling capabilities into Vision-Language-Action (VLA) models through action learning guided by reasoning.
Neutral-Reference Prompting for Vision–Language Models
Senmao Tian (Beijing Jiaotong University), Shunli Zhang (Beijing Jiaotong University)
ClassificationRecognitionDomain AdaptationComputational EfficiencyRepresentation LearningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a parameter-free plug-and-play prompting correction strategy called NeRP (Neutral-Reference Prompting), which utilizes neutral text prompts and reference images to evaluate and correct asymmetric confusion in transfer learning, thereby improving the recognition performance of unseen classes.
New Algorithms for Fully-Dynamic k-center with Outliers
Junyu Huang (Central South University), Qilong Feng (Central South University)
Anomaly DetectionOptimization
🎯 What it does: Proposed a hierarchical sampling framework for efficiently maintaining k-center clustering (with outliers) in fully dynamic environments, achieving an O(1) approximation and discarding only (1+ϵz) outliers.
New Bounds for Kernel Sums via Fast Spherical Embeddings
Tal Wagner (Tel Aviv University)
OptimizationComputational Efficiency
🎯 What it does: This paper proposes a new fast spherical embedding technique to improve the query time of high-dimensional kernel density estimation (KDE), and provides a new upper bound of ˜O(d+Δ+1/ε³) for Gaussian kernels;
New Wide-Net-Casting Jailbreak Attacks Risk Large Models
Qiuchi Xiang (Lancaster University), Jun Liu (Lancaster University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringGenerative Adversarial NetworkTextMultimodality
🎯 What it does: This paper studies the jailbreak scenario of 'wide-net-casting' and designs a multi-model collaborative attack method for this scenario.
Newton-coupled Dual-Teacher Semi-supervised Learning Framework
Hongyang He (University of Warwick), Victor Sanchez (University of Warwick)
ClassificationDomain AdaptationOptimizationKnowledge DistillationTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Propose a dual-teacher Newton-guided semi-supervised learning framework (TTN), which enhances the student model performance by fusing pseudo-labels from MAE and DINOv3 with Hessian, and using second-order optimization.
Next-Gen CAPTCHAs: Leveraging the Cognitive Gap for Scalable and Diverse GUI-Agent Defense
Jiacheng Liu (Mohamed bin Zayed University of Artificial Intelligence), Zhiqiang Shen (Mohamed bin Zayed University of Artificial Intelligence)
Safty and PrivacyTransformerLarge Language ModelAgentic AIPrompt EngineeringImageTextMultimodality
🎯 What it does: Propose a Next-Gen CAPTCHA framework, design 27 types of interactive challenges targeting the cognitive differences between GUI-Agent, and build a sustainable generation and verifiable evaluation platform.
NExT-Guard: Training-Free Streaming Safeguard without Token-Level Labels
Junfeng Fang (National University of Singapore), Xiang Wang (Shanghai Artificial Intelligence Laboratory)
Safty and PrivacyTransformerAuto EncoderContrastive LearningText
🎯 What it does: Propose the NEXT-GUARD framework, which utilizes Sparse Autoencoder (SAE) to extract implicit risk signals from pre-trained post-hoc safety models, achieving streaming safety protection without token-level annotation.
Next-Token Prediction and Regret Minimization
Mehryar Mohri (Google Research), Yifan Wu (Microsoft Research New England)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerTextSequential
🎯 What it does: This paper studies how to make next-word prediction models robust, enabling them to achieve low regret in online decision-making under adversarial decision environments;
NITP: Next Implicit Token Prediction for LLM Pre-training
Xiangdong Zhang (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Representation LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningText
🎯 What it does: Proposed the Next Implicit Token Prediction (NITP) objective, complementing the standard Next Token Prediction (NTP), enabling the model to learn the representation of the next token in a shallow semantic space, thereby reinforcing the geometric structure of hidden representations.
NL2Repo-Bench: Towards Long-Horizon Repository Generation Evaluation of Coding Agents
Jingzhe Ding (Bytedance Seed), Wenhao Huang (Bytedance Seed)
AI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose the NL2Repo-Bench benchmark, which generates complete installable Python repositories from natural language requirement documents from scratch, and rigorously verifies them using the original pytest test suite.
NNiT: Width-Agnostic Neural Network Generation with Structurally Aligned Weight Spaces
Jiwoo Kim (Duke University), Miroslav Pajic (Duke University)
Robotic IntelligenceReinforcement Learning from Human FeedbackNeural Architecture SearchConvolutional Neural NetworkGraph Neural NetworkTransformerMixture of ExpertsDiffusion modelScore-based ModelGenerative Adversarial NetworkImageTabular
🎯 What it does: Proposed a width-agnostic neural network generation framework called NNiT, which can directly generate effective weights on MLPs with arbitrary width and depth.
No Data? No Problem: Robust Vision-Tabular Learning with Missing Values
Marta Hasny (Technical University of Munich), Julia Schnabel (Technical University of Munich)
ClassificationRepresentation LearningData-Centric LearningTransformerContrastive LearningImageTabularBiomedical DataMagnetic Resonance ImagingElectronic Health Records
🎯 What it does: Propose the RoVTL framework, which maintains robustness in visual-table learning even when tables are missing to any degree (0%–100%).
No Free Lunch: Non-Asymptotic Analysis of Prediction-Powered Inference
Pranav Mani (Abridge AI), Michael Oberst (Johns Hopkins University)
Protein Structure PredictionContrastive LearningBiomedical DataBenchmark
🎯 What it does: Studied the true variance performance of prediction-driven inference (PPI++) for mean estimation under limited sample conditions, and provided an exact non-asymptotic expression.
No More K-means: Single-Stage Sparse Coding for Efficient Multi-Vector Retrieval
Lixuan Guo, Chenyu You (Stony Brook University)
RetrievalComputational EfficiencyRepresentation LearningTransformerAuto EncoderContrastive LearningText
🎯 What it does: Propose a single-stage sparse retrieval (SSR) method, which maps token embeddings into a high-dimensional sparse space via a sparse autoencoder, directly constructing a neuron-level inverted index, eliminating K-means clustering and multi-stage pruning;
No More, No Less: Least-Privilege Language Models
Paulius Rauba (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Designed and implemented a least-privilege language model framework and Nested Least-Privilege Networks (NLPN), dynamically restricting internal computations to adjust the model's capabilities on a per-request basis without retraining the model.
No Need to Train Your RDB Foundation Model
Linjie Xu (University of Hong Kong), David Wipf (University of Hong Kong)
ClassificationComputational EfficiencyData-Centric LearningGraph Neural NetworkTransformerPrompt EngineeringAuto EncoderContrastive LearningTabularBenchmark
🎯 What it does: This paper proposes an RDB foundation model framework that requires no additional training, using only vertical compression columns and a parameter-free JUICE encoder to work with existing single-table ICL models (such as TabPFN), achieving zero-training relational database prediction.
No Retraining at Edge: Efficient Resource-Aware Mixed-Precision Quantization via Federated Supernet Learning
Lianbo Ma (Northeastern University), Xingwei Wang (Northeastern University)
Federated LearningComputational EfficiencyRepresentation LearningConvolutional Neural NetworkTransformerMixture of ExpertsContrastive LearningImage
🎯 What it does: Train a shared mixed-precision quantized super network through federated learning, enabling edge devices to quickly extract suitable subnetworks under real-time resource constraints without requiring additional retraining;
Noise as a Natural Regularizer in Markov Decision Processes: Connecting Environmental Stochasticity and Policy Simplicity
Harry Chen (Massachusetts Institute of Technology), Ronald Parr (Duke University)
Reinforcement LearningTabularTime Series
🎯 What it does: Investigate the impact of environmental noise on the planning depth of Markov Decision Processes (MDPs), proving that solving noisy MDPs under various noise models is equivalent to solving noise-free MDPs with a smaller discount factor;
Noise-corrected GRPO: From Noisy Rewards to Unbiased Gradients
Omar El mansouri, Salem Lahlou (Mohamed bin Zayed University of Artificial Intelligence)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningText
🎯 What it does: Proposes the noise-corrected GRPO/Dr.GRPO methods, which utilize a Bernoulli noise model to denoise rewards and recover unbiased gradients, thereby enhancing the robustness of policy optimization in RLHF/RLVR environments.
Noise-Guided Transport: Imitation Learning from Random Priors
Lionel Blondé (University of Applied Sciences Western Switzerland), Alexandros Kalousis (University of Applied Sciences Western Switzerland)
Recurrent Neural NetworkTransformerReinforcement LearningScore-based ModelContrastive LearningTabularTime Series
🎯 What it does: Propose a low-sample scenario offline/online imitation learning method called Noise-Guided Transport (NGT), which constructs a reward function based on the prediction error of a random prior network, thereby achieving adversarial learning between expert and agent distributions.
Noise-Robust Density Estimation for Tabular Data Anomaly Detection
Dazhi Fu (Chinese University of Hong Kong), Jicong Fan (Hefei University of Technology)
Anomaly DetectionFlow-based ModelTabular
🎯 What it does: Propose a noise-robust density estimation method called NRDE, which uses Jacobian-regularized normalizing flows to separate clean data from noise sources, thereby performing discrimination based solely on the density of clean data in table data anomaly detection.
NoiseSDF2NoiseSDF: Learning Clean Neural Fields from Noisy Supervision
Tengkai Wang (Australian National University), Nick Barnes (Australian National University)
RestorationDiffusion modelScore-based ModelNeural Radiance FieldContrastive LearningPoint Cloud
🎯 What it does: Propose the NoiseSDF2NoiseSDF method, which performs unsupervised denoising of neural SDFs using noisy point cloud pairs to recover clean surfaces.
Noisy Pairwise-Comparison Random Search for Smooth Nonconvex Optimization
Taha EL BAKKALI EL KADI, Omar Saadi
OptimizationReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningTextTabular
🎯 What it does: This paper studies smooth non-convex optimization problems where only noisy comparative feedback is available, proposes the Noisy Comparative Random Search (NCRS) algorithm, and provides the query complexity of ε-first-order stopping points under low-dimensional active subspace structures;
Noisy-Channel Minimum Bayes Risk Decoding
Yusuke Sakai (Nara Institute of Science and Technology), Taro Watanabe (Nara Institute of Science and Technology)
GenerationData SynthesisTransformerReinforcement LearningContrastive LearningTextMultimodalityBenchmark
🎯 What it does: Studied the MBR decoding method based on noisy channel decomposition, explaining different decoding variants by decomposing MBR into four probability channels.
Noisy-Space Policy Gradient for Diffusion Policies in Offline Reinforcement Learning
Mahmoud Selim (TRATON CV AB), Karl Henrik Johansson
Reinforcement LearningDiffusion modelScore-based ModelImageVideoTabularBenchmark
🎯 What it does: Propose a value-based optimization framework that constructs noise space action value functions and noise space policy gradients for diffusion policies in offline reinforcement learning.
NOMAD: Lifelong Trajectory Planning via Non-Parametric Bayesian Memory-Adaptive Diffusion Experts
Yixian Chen (Foshan University), Yuhuan Lu (Macao Polytechnic University)
Autonomous DrivingOptimizationTransformerMixture of ExpertsDiffusion modelScore-based ModelGenerative Adversarial NetworkPoint CloudTabularBenchmark
🎯 What it does: Built a lifelong trajectory planning framework called NOMAD, which utilizes nonparametric Bayesian memory and diffusion experts to achieve continuous adaptation to long-tail scenarios.
Non-Adversarial Imitation Learning Provably Free of Compounding Errors: The Value Flow Mechanism
Tian Xu (Nanjing University), Yang Yu (Nanjing University)
Reinforcement LearningTabularTime SeriesSequentialBenchmark
🎯 What it does: This paper proposes a new non-adversarial imitation learning method called Dual Q-DM, aiming to solve the composite error problem in existing methods and achieve better generalization ability through a value flow mechanism.
Non-Euclidean Gradient Descent Operates at the Edge of Stability
Rustem Islamov (University of Basel), Robert M. Gower (Flatiron Institute)
OptimizationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelFlow-based ModelRectified FlowAuto EncoderGenerative Adversarial NetworkContrastive LearningImageText
🎯 What it does: This paper investigates the Edge of Stability (EoS) phenomenon in gradient descent (GD) under non-Euclidean norms, introducing the concepts of directional smoothness and generalized sharpness, which are used to explain and quantify EoS behaviors across different geometries.
Non-Monotonic Autoregressive Sequence Model
Tianyi Ma (University of Notre Dame), Yanfang Ye (University of Notre Dame)
GenerationOptimizationComputational EfficiencyAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextSequential
🎯 What it does: Proposes N-MARS, a non-monotonic autoregressive sequence model that utilizes a special <UNDO> token to enable immediate corrections during the generation process.
Non-Parametric Optimization for Scalable Learning in Stochastic Decision Problems
Mohsen Amidzadeh (Aalto University), Mario Di Francesco (Aalto University)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTabularTime SeriesStochastic Differential Equation
🎯 What it does: This study investigates time-varying stochastic optimization (TV-SO), proposes nonparametric optimality conditions, and designs a scalable deep learning algorithm (SPF) to solve stochastic decision problems with distribution drift.
Non-Parametric Probabilistic Robustness: A Conservative Risk Estimator under Unknown Perturbation Distributions
Zheng Wang (University of Warwick), Xingyu Zhao (University of Warwick)
ClassificationAnomaly DetectionAdversarial AttackRobotic IntelligenceConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Proposes a nonparametric probabilistic robustness evaluation framework, NPPR, which learns the most conservative perturbation distribution to estimate robustness under unknown perturbation distributions.
Non-Parametric Structural Priors for Geometry Theorem Prediction
Junbo Zhao (Beijing Normal University), Hua Huang (Beijing Normal University)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Achieve multi-step theorem prediction in geometric reasoning by using large language models (LLMs) combined with retrieval-enhanced theorem priority graphs (TPG) in an untrained manner.
Non-Stationary Online Structured Prediction with Surrogate Losses
Shinsaku Sakaue (CyberAgent), Yuzhou Cao (Nanyang Technological University)
OptimizationMeta LearningReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningContrastive LearningTabularTime SeriesSequential
🎯 What it does: Studied non-stationary online structured prediction, proposed an upper bound of 'small-surrogate-loss + path-length', and achieved better target loss control through Polyak-style learning rate
Non-Uniform Noise-to-Signal Ratio in the REINFORCE Policy-Gradient Estimator
Haoyu Han (Harvard University), Heng Yang (Harvard University)
OptimizationReinforcement LearningContrastive LearningGaussian SplattingTabularTime SeriesSequentialStochastic Differential Equation
🎯 What it does: Analyze the noise-to-signal ratio (NSR) of the REINFORCE policy gradient estimator, and provide exact expressions or upper bounds under linear, polynomial, and general nonlinear dynamics, demonstrating the non-uniform distribution of NSR in the parameter space and explaining its evolution during optimization and the mechanisms leading to training instability.
Nonconvex Low-Rank Tensor Representation with Deep Priors for Multiview Subspace Clustering
Yao Fu (Southwest University), Zhi Wang (Southwest University)
OptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningAuto EncoderContrastive LearningImageMultimodalityTabularBenchmark
🎯 What it does: Proposed a multi-view subspace clustering model called NRDN-MvSC, which combines a non-convex low-rank tensor representation with a deep prior.
Nonlinear Covariate Balance in Experimental Design
Qing Chen (Rutgers University), Peng Zhang (Rutgers University)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyData-Centric LearningAuto EncoderContrastive LearningTabularBenchmark
🎯 What it does: Proposed an experimental design framework based on Gram-Schmidt Walk (GSW), which achieves balance of nonlinear covariate structures in random assignment by balancing covariates in a nonlinear feature space.
Nonparametric Data Attribution for Diffusion Models
Yutian Zhao (Sea AI Lab), Min Lin (Sea AI Lab)
GenerationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkDiffusion modelScore-based ModelAuto EncoderContrastive LearningImage
🎯 What it does: Proposed a non-parametric data attribution method (NDA), which evaluates the impact of each training sample on the generation results of diffusion models by quantifying the patch-level similarity between generated images and training data, without requiring access to model gradients or retraining.
Nonparametric Distribution Regression Re-calibration
Ádám Jung (HUN-REN SZTAKI), Andras A Benczur
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTabularBenchmark
🎯 What it does: Proposed a nonparametric recalibration algorithm based on conditional kernel mean embedding, aiming to address the calibration issue between the predicted distribution and the true empirical uncertainty in probabilistic regression.
Nonparametric LLM Evaluation from Preference Data
Dennis Frauen (LMU Munich), Stefan Feuerriegel (LMU Munich)
Recommendation SystemData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelDiffusion modelScore-based ModelFlow-based ModelRectified FlowNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes a non-parametric framework DMLRANK, which compares and ranks LLMs using generalized average ranking scores (GARS) based on preference data, and achieves unbiased estimation and confidence intervals through double machine learning.
NonZero: Interaction-Guided Exploration for Multi-Agent Monte Carlo Tree Search
Sizhe Tang (George Washington University), Tian Lan (George Washington University)
TransformerReinforcement LearningMixture of ExpertsWorld ModelOptical FlowTabularSequentialBenchmark
🎯 What it does: Propose the NONZERO framework, which achieves scalable search in the exponential joint action space of multi-agent MCTS through interaction-guided candidate exploration.
Norm$\times$Direction: Restoring the Missing Query Norm in Vision Linear Attention
Weikang Meng (Harbin Institute of Technology), Zheng Zhang (Harbin Institute of Technology)
ClassificationRestorationSegmentationSuper ResolutionComputational EfficiencyRepresentation LearningTransformerVision Language ModelDiffusion modelContrastive LearningImageMultimodality
🎯 What it does: Propose the NaLaFormer linear attention mechanism, restoring the norm information of query vectors and adopting cosine directional similarity to maintain non-negativity, thereby enhancing the expressiveness and efficiency in visual tasks.
Normality Calibration in Semi-supervised Graph Anomaly Detection
Guolei Zeng (University of Oxford), Guansong Pang (Singapore Management University)
Anomaly DetectionKnowledge DistillationGraph Neural NetworkAuto EncoderContrastive LearningGraph
🎯 What it does: Propose the GraphNC framework, which uses a pre-trained teacher model to perform normality calibration for semi-supervised graph anomaly detection in the score space and representation space.
Normalization Equivariance for Arbitrary Backbones, with Application to Image Denoising
Youssef Saied (University of Geneva), François Fleuret (University of Geneva)
RestorationConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningImageBenchmark
🎯 What it does: Propose a parameter-free wrapper (Wrapped Normalization Equivariance, WNE), which achieves normalization equivariance (NE) for any network by normalizing the input, invoking any backbone, and then denormalizing, thereby enhancing robustness to mismatched noise levels in blind image denoising tasks.
Normalization-equivariant Diffusion Models: Learning Posterior Samplers From Noisy And Partial Measurements
Brett Levac (University of Texas at Austin), Julián Tachella (ENS de Lyon)
Image TranslationRestorationGenerationData SynthesisSuper ResolutionDiffusion modelScore-based ModelAuto EncoderImageBiomedical DataMagnetic Resonance ImagingStochastic Differential Equation
🎯 What it does: Proposes a framework based on the normalized equivariant diffusion model (NE-Diffusion), which can train a posterior sampler using only noisy measurement data from a single denoising or linear inverse problem measurement, without requiring clean images, achieving image denoising, demosaicking, missing data filling, and low-field MRI reconstruction;
Normalized Rewards for Preference Optimization
Shawn Im (University Of Wisconsin Madison), Katherine Metcalf (Apple)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: This paper introduces a regularization term to maintain the total probability of observed responses, addressing the likelihood shift problem that occurs in direct alignment algorithms (such as DPO and SimPO), and improves upon these two algorithms.
Normalizing Flows with Iterative Denoising
Tianrong Chen (Apple), Shuangfei Zhai (Apple)
GenerationTransformerDiffusion modelFlow-based ModelImageOrdinary Differential Equation
🎯 What it does: Propose an iterative denoising regularized flow model called iTARFlow, which utilizes Transformer normalization flows trained with multiple noise levels. During inference, it first generates high-noise samples and then recovers images through iterative denoising.
NorMuon: Making Muon more efficient and scalable
Zichong Li (Georgia Institute of Technology), Tuo Zhao (Georgia Institute of Technology)
OptimizationComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed the NorMuon optimizer for efficient training of large-scale language models.
Not All Answers Are Contextually Persuadable: Inference Dynamics in Large Language Models under Contextual Influence
Zongye Hu (Arizona State University), Ziyi Huang (Arizona State University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper proposes and verifies a theoretical framework to quantify the internal reasoning dynamics of large language models under repeated contexts (such as multiple repetitions of the same assertion), and provides a proof that the reasoning state converges to a stable interval; meanwhile, based on this framework, a representation layer estimation method requiring only a single forward inference is developed to predict the final change in answer tendency under infinite repetition.
Not All Frequencies Are Equal: Energy-Adaptive Diffusion for Time Series Forecasting
Zining Qin (Beijing Normal-Hong Kong Baptist University), Weijia Jia (Beijing Normal University-Zhuhai)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelTabularTime SeriesBenchmarkFinance RelatedPhysics RelatedStochastic Differential Equation
🎯 What it does: Propose EADIFF, an energy-adaptive diffusion model based on the waveform domain, for time series forecasting
Not All Invariants Are Equal: Curating Training Data to Accelerate Program Verification with SLMs
Ido Pinto (Hebrew University of Jerusalem), Guy Katz (Hebrew University of Jerusalem)
Computational EfficiencyData-Centric LearningAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextSequentialRetrieval-Augmented Generation
🎯 What it does: Proposed the WONDA data cleaning pipeline and fine-tuned small language models to generate higher quality loop invariants, significantly accelerating program verification.
Not All Prefills Are Equal: PPD Disaggregation for Multi-turn LLM Serving
Zongze Li (University of Chicago), Ce Zhang (University of Chicago)
OptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper studies the Prefill-Decode (PD) splitting architecture in multi-round LLM inference services, identifying inefficiencies in KV transmission and recomputation in multi-round dialogue scenarios. It proposes a PPD scheme based on dynamic routing, allowing subsequent requests to complete Append-Prefill locally at the Decode node, thus reducing KV transmission and prefill computation.
Numina-Lean-Agent: An Open and General Agentic Reasoning System for Formal Mathematics
Junqi Liu (University of Chinese Academy of Sciences), Wenda Li (University of Edinburgh)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a formal mathematical reasoning framework called Numina-Lean-Agent based on a general-purpose coding agent, achieving a perfect score of 12/12 in the Putnam 2025 competition, and collaboratively formalizing a harmonic analysis paper with mathematicians using Lean.
OBCache: Optimal Brain KV Cache Pruning for Efficient Long-Context LLM Inference
Yuzhe Gu (University of Pennsylvania), Enmao Diao (DreamSoul)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Studied the KV cache pruning problem in long-context reasoning of LLMs, proposing the OBCache framework, which treats cache eviction as hierarchical structured pruning and provides token importance scores based on output perturbation.
Object-level Semantic and Spatial Distillation for Open Vocabulary Detection
Zitong Li (China University of Geosciences), Dapeng Luo (China University of Geosciences)
Object DetectionKnowledge DistillationConvolutional Neural NetworkTransformerVision Language ModelAuto EncoderContrastive LearningImage
🎯 What it does: Proposes a two-stage semantic and spatial distillation framework called OSSD for open-vocabulary object detection, decoupling semantic and spatial feature learning to significantly improve detection performance.
ObjEmbed: Towards Universal Multimodal Object Embeddings
Shenghao Fu (Sun Yat-sen University), Wei-Shi Zheng (Sun Yat-sen University)
Object DetectionRetrievalRepresentation LearningTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A large-scale multi-modal language model (LLM)-based object-level embedding model, ObjEmbed, was constructed, which can simultaneously generate semantic embeddings and localization information (IoU embeddings) for all image objects in a single forward inference, and is compatible with multiple tasks such as object detection, referring expression understanding, and local and global image retrieval.
OBJVanish: Prompt-Driven Generation of Physically Realizable 3D LiDAR-Invisible Objects
Bing Li (Nanyang Technological University), Qing Guo (Nankai University)
Autonomous DrivingAdversarial AttackTransformerPrompt EngineeringDiffusion modelGaussian SplattingTextPoint Cloud
🎯 What it does: This paper proposes a text-prompt-based 3D object generation framework called OBJVanish, which can generate objects invisible to LiDAR, thereby enabling physically deployable adversarial attacks.
Obliviate: Efficient Unlearning in Recommender Systems
Tushar Prakash (Sony Research India), Narayan Chaturvedi
Recommendation SystemComputational EfficiencyKnowledge DistillationGraph Neural NetworkContrastive LearningGraphTabular
🎯 What it does: An efficient machine forgetting framework called Obliviate is proposed for recommendation systems, which can quickly remove the influence of user interaction data without complete retraining.
OC-space: a Unifying Perspective on Verification of Tree Ensembles
Timo Martens (KU Leuven), Jesse Davis (KU Leuven)
Explainability and InterpretabilityComputational EfficiencyTabularBenchmark
🎯 What it does: This paper proposes a unified verification framework based on the output configuration space (OC-space) of tree ensemble models, which determines whether a model satisfies a given property by searching over all possible leaf combinations.
OcclusionFormer: Arranging Z-Order for Layout-Grounded Image Generation
Ziye Li (Fudan University), Henghui Ding (Fudan University)
GenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelFlow-based ModelImageMultimodality
🎯 What it does: Designed and implemented OcclusionFormer, a layout-guided image generation framework based on diffusion Transformer, explicitly modeling Z-order to address occlusion issues caused by object overlaps.
OCNR: Stabilizing Self-Play by Mitigating Iteration-Collapse With One-Class Novelty Rewards
Seungyoo Lee (KAIST), Juho Lee (KAIST)
Anomaly DetectionOptimizationRepresentation LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: Propose a One-Class Novelty Reward (OCNR) mechanism, which uses a 'seen detector' during self-play to determine whether a problem is a previously encountered training instance, and penalizes the repetition of generated tasks, thereby suppressing iteration collapse and stabilizing the later stages of LLM training.
Off-Policy Evaluation for Missingness-Aware Policies in MDPs with Rewards Missing Not at Random
Ziheng Wei (University of Michigan), Rui Miao (University of Texas)
Reinforcement LearningAuto EncoderContrastive LearningTabularBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes an off-policy evaluation (OPE) method for offline reinforcement learning under the missing not at random (MNAR) scenario, which can estimate the value of the target policy using log data with only partial reward observations;
Off-Policy Evaluation with Strategic Agents via Local Disclosure
Kiet Q. H. Vo (CISPA Helmholtz Center for Information Security), Krikamol Muandet (CISPA Helmholtz Center for Information Security)
Federated LearningExplainability and InterpretabilityData-Centric LearningReinforcement LearningTabularFinance Related
🎯 What it does: Studied the problem of offline policy evaluation caused by agents strategically modifying features, utilizing local information disclosure to obtain pre-policy features, constructing a consistent double robust estimator, and verifying it on synthetic data and German Credit data.
Off-Policy Learning in Large Action Spaces: Optimization Matters More Than Estimation
Imad Aouali (Criteo AI Lab), Otmane Sakhi (Criteo AI Lab)
Recommendation SystemOptimizationReinforcement LearningContrastive LearningTabular
🎯 What it does: This paper addresses the problem of offline contextual multi-armed bandits with large action spaces, proposing a policy learning framework centered on optimization difficulty, emphasizing the importance of optimization stability.
Offline Multi-agent Continual Cooperation via Skill Partition and Reuse
Yuchen Xiao (Nanjing University), Yang Yu (Nanjing University)
TransformerReinforcement LearningAuto EncoderContrastive LearningTabularSequentialBenchmark
🎯 What it does: Proposed the COMAD framework, which enables continuous skill discovery and transfer in offline multi-agent continual cooperation through skill segmentation and reuse.
Offline Multi-Agent Reinforcement Learning via Sequential Score Decomposition
Dan Qiao (Chinese University of Hong Kong), Baoxiang Wang (Chinese University of Hong Kong)
TransformerReinforcement LearningDiffusion modelScore-based ModelSequential
🎯 What it does: Propose an offline multi-agent reinforcement learning framework called OMSD, which utilizes sequential decomposition of behavioral policies and employs diffusion models to estimate conditional scores for behavioral regularization, encouraging agents to maintain coordination within offline data.
Offline Preference Optimization for Rectified Flow with Noise-Tracked Pairs
Yunhong Lu (Zhejiang University), Min Zhang (Zhejiang University)
GenerationOptimizationReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningDiffusion modelScore-based ModelFlow-based ModelRectified FlowImageTextMultimodality
🎯 What it does: Propose an offline preference optimization framework called PNAPO for rectified flow text-to-image models, which refines the alignment of generated trajectories by utilizing prior noise recorded during training.
Offline Reinforcement Learning of High-Quality Behaviors Under Robust Style Alignment
Mathieu Petitbois (Ubisoft La Forge), Sylvain Lamprier (Université d'Angers)
Reinforcement LearningTabularTime SeriesSequential
🎯 What it does: This paper proposes a unified definition of behavioral style and develops an offline reinforcement learning framework called SCIQL, aiming to achieve a high-quality balance between style and task performance.
Offline Reinforcement Learning with Generative Trajectory Policies
Xinsong Feng (William & Mary), Haipeng Chen (William & Mary)
Reinforcement LearningDiffusion modelScore-based ModelFlow-based ModelTabularTime SeriesBenchmarkOrdinary Differential Equation
🎯 What it does: This paper proposes Generative Trajectory Policies (GTP), a general generative policy based on continuous-time ODEs for offline reinforcement learning.
Offline Reinforcement Learning with Universal Horizon Models
Hojun Chung (Seoul National University), Songhwai Oh (Seoul National University)
Reinforcement LearningFlow-based ModelWorld ModelTabularTime SeriesSequentialBenchmark
🎯 What it does: Propose a universal horizon model (UHM) that can directly sample future states at arbitrary time steps and build a scalable offline reinforcement learning framework;
Offline Two-Player Zero-Sum Markov Games with KL Regularization
Claire Chen (California Institute of Technology), Nan Jiang (University of Illinois Urbana Champaign)
OptimizationReinforcement LearningContrastive Learning
🎯 What it does: Proposes an offline two-player zero-sum Markov game learning framework based on KL regularization. First, it proves that rapid convergence can be achieved under the single-sided coverage condition through the ROSE theoretical framework; subsequently, it designs a practical SOS-MD algorithm (Sequential Offline Self-Play Mirror Descent) to achieve approximate equilibrium solving.
OGPO: Sample Efficient Full-Finetuning of Generative Control Policies
Sarvesh Patil (Carnegie Mellon University), Max Simchowitz (Carnegie Mellon University)
Robotic IntelligenceTransformerSupervised Fine-TuningReinforcement LearningDiffusion modelFlow-based ModelMultimodality
🎯 What it does: Propose OGPO, a sample-efficient full-policy fine-tuning method specifically designed for generative control policies (GCP) that use diffusion or flow models;
Olaf-World: Orienting Latent Actions for Video World Modeling
Yuxin Jiang (Show Lab, National University of Singapore), Mike Zheng Shou (Show Lab, National University of Singapore)
Domain AdaptationRepresentation LearningReinforcement Learning from Human FeedbackTransformerVision-Language-Action ModelDiffusion modelAuto EncoderContrastive LearningWorld ModelImageVideo
🎯 What it does: Proposes an unsupervised latent action learning and world model pre-training framework called Olaf-World, which can learn action representations transferable across domains from unlabeled videos, and use these representations to achieve zero-shot action transfer and efficient task adaptation.
Old Habits Die Hard: How Conversational History Geometrically Traps LLMs
Adi Simhi (Technion Israel Institute of Technology), Shay B Cohen (University of Edinburgh)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This study proposes the HISTORY-ECHOES framework, which explores how large language models (LLMs) tend to generate specific responses under the influence of dialogue history, especially how they maintain consistent behavior in multi-turn conversations.
OLion: Approaching the Hadamard Ideal by Intersecting Spectral and L inf Implicit Biases
Zixiao Wang (Peking University), Huishuai Zhang (Peking University)
OptimizationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningMixture of ExpertsImageText
🎯 What it does: Proposed a new optimizer called OLion (Orthogonal Lion), which first performs Newton-Schulz orthogonalization on the gradient, then takes the sign of the orthogonalized direction, combining Muon's spectral structure control and Lion's ℓ∞ coordinate control, achieving Hadamard idealized updates for matrix parameters;
Olivia: Harmonizing Time Series Foundation Models with Power Spectral Density
Jingru Fei (Beijing Institute of Technology), Wei Fan (University of Auckland)
Computational EfficiencyRepresentation LearningTransformerMixture of ExpertsAuto EncoderContrastive LearningTime Series
🎯 What it does: Propose the Olivia model, which combines Harmonizer and HarmonicAttention for spectral domain alignment and efficient attention in time series foundational models.
Olmix: A Framework for Data Mixing Throughout LM Development
Mayee F Chen (Allen Institute for AI), Kyle Lo (Allen Institute for AI)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelMixture of ExpertsTextReview/Survey Paper
🎯 What it does: This study proposes the OLMIX framework to systematically address the data mixing process in language model training, and conducts in-depth analysis and solutions for the two major challenges: configuration of mixing methods and the evolution of domain sets over time.
OMAC: A Holistic Optimization Framework for LLM-Based Multi-Agent Collaboration
Shijun Li (University of Texas at Austin), Joydeep Ghosh (University of Texas at Austin)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringContrastive LearningText
🎯 What it does: Proposed the OMAC framework for holistic optimization of LLM-driven multi-agent systems in multi-step collaborative scenarios;
Omitted Variable Bias in Language Models Under Distribution Shift
Victoria Lin (Carnegie Mellon University), Eli Ben-Michael (Carnegie Mellon University)
Domain AdaptationOptimizationExplainability and InterpretabilityTransformerLarge Language ModelTextTabularTime SeriesSequential
🎯 What it does: The study investigates omitted variable bias (OVB) in language models under distributional shift, and proposes a framework that maps the strength of omitted variables to the bounds of the model's worst-case generalization performance.
Omni-Diffusion: Unified Multimodal Understanding and Generation with Masked Discrete Diffusion
lijiang Li, Chaoyou Fu (Nanjing University)
GenerationData SynthesisRepresentation LearningTransformerLarge Language ModelVision Language ModelDiffusion modelContrastive LearningImageTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: Designed and implemented Omni-Diffusion, a multi-modal language model based on mask-based discrete diffusion models, capable of unified understanding and generation of text, images, and speech.
Omni-fMRI: A Universal Atlas-Free fMRI Foundation Model
Mo Wang (Southern University of Science and Technology), Quanying Liu (Southern University of Science and Technology)
ClassificationRepresentation LearningTransformerAuto EncoderContrastive LearningBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This study proposes Omni-fMRI, a voxel-level fMRI foundational model that does not require predefined brain region parcellation, capable of directly performing self-supervised pre-training on full-brain 4D BOLD signals;
Omni-Perception Policy Optimization for Multimodal Emotion Reasoning
Zhiyuan Han (University of Science and Technology of China), Xun Yang (University of Science and Technology of China)
OptimizationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-ThoughtAudio
🎯 What it does: Proposed a framework called OPPO based on reinforcement learning, which explicitly optimizes full-modal perception in multimodal emotion reasoning, ensuring that the reasoning process fully utilizes visual, auditory, and textual cues.
OmniAID: Decoupling Semantic and Artifacts for Universal AI-Generated Image Detection in the Wild
Yuncheng Guo (Shanghai Artificial Intelligence Laboratory), Weijia Li (Tsinghua University)
Anomaly DetectionTransformerPrompt EngineeringMixture of ExpertsDiffusion modelContrastive LearningImageMultimodality
🎯 What it does: Built a hybrid expert model that can decouple semantic defects from generator traces for detecting AI-generated images